[Clang][Sema] Fix lookup of dependent operator= named by using-declaration

[sdkrystian]

Fixes this bug caused by #90152.

Will add tests shortly.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Krystian Stasiowski (sdkrystian)

Changes

Fixes this bug caused by #90152.

Will add tests shortly.

---

Full diff: https://github.com/llvm/llvm-project/pull/91503.diff

1 Files Affected:

• (modified) clang/lib/Sema/SemaDeclCXX.cpp (+12-5)

diff --git a/clang/lib/Sema/SemaDeclCXX.cpp b/clang/lib/Sema/SemaDeclCXX.cpp
index 157d42c09cfcd..91c83564b567e 100644
--- a/clang/lib/Sema/SemaDeclCXX.cpp
+++ b/clang/lib/Sema/SemaDeclCXX.cpp
@@ -12963,14 +12963,15 @@ NamedDecl *Sema::BuildUsingDeclaration(
     return nullptr;
   }
 
-  DeclContext *LookupContext = computeDeclContext(SS);
   NestedNameSpecifierLoc QualifierLoc = SS.getWithLocInContext(Context);
-  if (!LookupContext || EllipsisLoc.isValid()) {
-    NamedDecl *D;
+  DeclContext *LookupContext = computeDeclContext(SS);
+
+  auto BuildDependent = [&] {
+    NamedDecl *D = nullptr;
     // Dependent scope, or an unexpanded pack
     if (!LookupContext && CheckUsingDeclQualifier(UsingLoc, HasTypenameKeyword,
                                                   SS, NameInfo, IdentLoc))
-      return nullptr;
+      return D;
 
     if (HasTypenameKeyword) {
       // FIXME: not all declaration name kinds are legal here
@@ -12987,7 +12988,7 @@ NamedDecl *Sema::BuildUsingDeclaration(
     CurContext->addDecl(D);
     ProcessDeclAttributeList(S, D, AttrList);
     return D;
-  }
+  };
 
   auto Build = [&](bool Invalid) {
     UsingDecl *UD =
@@ -13002,6 +13003,9 @@ NamedDecl *Sema::BuildUsingDeclaration(
   auto BuildInvalid = [&]{ return Build(true); };
   auto BuildValid = [&]{ return Build(false); };
 
+  if (!LookupContext || EllipsisLoc.isValid())
+    return BuildDependent();
+
   if (RequireCompleteDeclContext(SS, LookupContext))
     return BuildInvalid();
 
@@ -13024,6 +13028,9 @@ NamedDecl *Sema::BuildUsingDeclaration(
 
   LookupQualifiedName(R, LookupContext);
 
+  if (R.wasNotFoundInCurrentInstantiation())
+    return BuildDependent();
+
   // Validate the context, now we have a lookup
   if (CheckUsingDeclQualifier(UsingLoc, HasTypenameKeyword, SS, NameInfo,
                               IdentLoc, &R))

[erichkeane]

Did the FIXME ever get a core issue? Perhaps we need to run that down a bit? @Endilll or @shafik , any chance the 'FIXME" that we just 'fixed' looks familiar?

[Endilll]

This rings a bell. I've put something similar in our DR tests, probably in a PR about complete-class context. I'm running out of time today, but I'll look into this tomorrow if Shafik doesn't beat me to it.

[sdkrystian]

The FIXME wasn't fixed, it's just broken in a different way that's consistent with other "current instantiation broken-ness" related things. I'm going to submit a patch (probably this week) which diagnoses cases where the current instantiation is used in a context which requires a complete type.

[erichkeane]

Ah, this is githubs fault :) I mean the OLD 'FIXME' that you removed. I was wondering if it has a CWG issue.

[Endilll]

There are somewhat related CWG issues 45, 374, and 500, but I can't find one arguing that base specifiers should become a complete-class context. http://eel.is/c++draft/class#mem.general-8 seems quite clear on this.

[sdkrystian]

Closing this for now; I'll return to this once we correctly handle dependent operator=.

[pawosm-arm]

The number of C++ workloads failing to build after #90152 is rather impressive, even this PR does not help. Is there a plan to address that?

[sdkrystian]

@pawosm-arm Could you provide some examples? Barring an issue with operator= which has been addressed, builds should only fail in the presence of accesses to non-existent members. Addressing such issues is the responsibility of users.

[pawosm-arm]

I'm sorry, I didn't notice that there is more discussion after #90152 covering also my issue at hand. The obvious advice I can get from those is the 'this is not valid c++, fix your code' suggestion. Trouble starts when there is a lot of it to fix, even if it mostly involve deletion of the functions that aren't instantiated.

NB, suggesting use of -fdelayed-template-parsing isn't always working, unless we start using something like -fdelayed-template-parsing -Wno-error=delayed-template-parsing-in-cxx20